Scribd
Upload
32 views17 pages

Weak Lower Semi Continuity

Broad explanation and properties of weak lower semi continuity

Uploaded by

prasanjit samantaray
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
32 views17 pages

Weak Lower Semi Continuity

Broad explanation and properties of weak lower semi continuity

Uploaded by

prasanjit samantaray
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF or read online on Scribd
You are on page 1/ 17
Quasi-Convex Functionals 141 In the theorem just proved the functions u(x) (= u) and z(z) were completely independent, so that it does not apply to the situation we are most interesteding, namely that of functionals of the type F(u,2) = [Fe u,Du)dz, (64) that is when z = Du. And, in fact, the functions z, and z involved in the proof of the theorem were not the gradients of the corresponding uj, and u, as it should be if we want to find necessary conditions for the semicontinuity of the functional (5.4). Of course, this restricts considerably the choice of the possible sequences zy + 2, 80 that the convexity in z, which we have shown to be necessary in the general case, might not be so essential in the new situation. In this case, if we continue to assume that the functions z, = Duy converge in the weak* topology of L®, so that ||Dunlloo < M, it will not be restrictive to assume that the sequence uj converges uniformly, since any other weaker convergence would reduce to that by the theorem of ASCOLI-ARZELA. In conclusion, we may assume that the functional F is LSC with respect to convergence in the weak* topology of W49°(9) = Lip (), or else, which is essentially equivalent, with respect to uniform convergence of bounded sequences of Lip () (L-convergence): Definition 5.1 A sequence un € Lip (®) és said to be L-convergent to a function u if un + u uniformly in Q and if there exists a constant M such that for every integer h fualo = |Dua(z)| 1. This being established, we shall consider first the special case in which the function F depends only on z. Theorem 5.2 Let Q be a bounded open set of R®, let F(z) be a conti- nuous function in R"’, and assume that the functional Flu) = f F(Du(z)) de 142 Direct Methods in the Calculus of Variations is lower semicontinuous with respect to L-convergence. Then, for every function p € Ch(Q) and for every z ER" we have F(20)|0| < f Feo + Do(a)) de. (5.6) Proof. Let Q be a cube containing Q, and that modulo a homothety we can suppose to be the unit cube [0, 1]". We can extend the function ¢ first setting it to be equal to zero in Q — Q, and then extending it periodically in R": sftn})s where with {a} we denote the fractionary part of a. For any integer h we set o(aay---12n) = v({a1}, en(2) = jo(hz), un(z) = (20,2) + Pn(z). It is clear that un 4 (zo,z), and hence by the semicontinuity assumption: [QLF (20) < lim int [ _ imint | = De(he), we get, after a change of variables, IQUF(20) < tim inf A=" f Pert Dolw)dy, where hQ = (0, h]". On the other hand y has period one, and hence also F(z9 + Dy) has the same period; it follows that. J Feat Dota) ay = f, F(2o + Doty) dy and hence QIF (20) < f F (20 + Doly)) dy. We can now deal with the general case. Quasi-Conver Functionals 143 Theorem 5.3 Let Q be a bounded open set of R", let F(x,u,z) be a continuous function in Q x R x R™, and assume that the functional F (ua) = [Pl u(e), Du(a) de is lower semicontinuous with respect to L-convergence. Then, for every function g € C}(Q) and for every x9 € 9, up € RX and z € R™ we have F(2o, uo, 0)|| < if F (20,40, 20+ Dy(2)) de. (5.7) Proof. As above, we can suppose that { be contained in the unit cube Q, and that ¢ is extended periodically in R”. Let xo € 9, and let W bea cube contained in 2, with center zo and side t. We define u(x) = uo + (zo, — 20), en(z) = ie (), ua(z) = u(x) + Ga(e), where : paz) ifcoew, Pr(z) = A 0 ifceQ-W. Since ya(x) = 0 in a neighborhood of OW, the function Gp belongs to cu). To get an estimate for F(u,,W), we decompose W into h” equal cubes W; of centers 23. We have Flu, W) = I, F(a,u+ en, 20+ Don) de = SI, F (sm(eds20 + De (=) dz . + © [Pet encoder) ~ F(ei,u(xi), 20 + Depn)) de = Ant By. 144 Direct Methods in the Calculus of Variations For what concerns the term B, we remark that the derivatives of yn are equibounded, and therefore the arguments of the function F remain confined in a compact set © C 9x RN x R°". The function F is then uniformly continuous in ©, and hence By —+ 0 when h -+ 0. On the other hand h(a; — zo)/t is a vector with integral components, and therefore by the periodicity of y we have he Sf F (seule)20 + Do A ) ohn, which, after a change of variables y = h(a — 2;)/t, becomes Pe n= 3 (f)" f Plewwteds20+ Dot) a. 1 We write for simplicity o(a) = [ F(a,u(e), 29 + Do(y)) dy. The function g is continuous, and we have i An= oles) i=l and hence tin An= ff de f Fle,u(a) 20 + Dely)) ay. ee ale On the other hand the sequence uy L-converges to u, so that the lower semicontinuity of F gives lim An = lim Flun,W) > F(u,W) = f F(2,u(2), 2) de iim, An = Jim, fe and the conclusion is obtained dividing by |W| and letting the side t of W go to zero. o ‘As we shall see, condition (5.7) will play a major role not only in the semicontinuity theory, but also in the regularity of minima. We are then induced to formulate the following definition, Definition 5.2 The functional Fu) = [Fewpy de Quasi-Convez Functionals 145, is called quasi-convex if for every (20, Uo, 20) € 2x RN x RY and for every y € C(O, RN) we have F(x, 40, 20) S ff Fl2o, oy 20-+ Dele)) de. (5.8) Geometrically, quasi-convexityindexquasi-convexity means that the linear functions minimize the “frozen” functional Fu.) =f Fle0, 1, Du(a)) de. ln Remark 5.1 We remark that if (5.8) holds for an open set ©, it holds also for every open set A CR", Actually, we can choose t and @ in such a way that Ay ={ceER™:te+ZEA}CQ. If supp (n) CA, the function g(x) = t~!n(y) =: t2y(tx + &) belongs to C4(Q), and hence f Pleo, vor20 + Dily)) dy = ff Fleo,uoy20 + Dole) de i fi = {| Plow nors0 + Dele) de . Feo, ¥0,20)!0—Ai)} 2 F (xo, 40,20) As a consequence we can speak of quasi-convex functions: a function F being quasi-conver if (5.8) holds in some open set 2c R™. a It is easily seen that quasi-convexity is strictly weaker than convexity. In fact, if F(,u, z) is a continuous function, convex in z, there holds F (20, to, zo + Dy) > F(x0, uo, 20) + (Ao, Dy) for some Ay € R™. Integrating over 2, J Pleo... 29+ Dp) de > Fla, 10,20) 146 Direct Methods in the Calculus of Variations since [0o.Derdz =o. A It follows that any convex function F is quasi-convex. The converse is not true, since the function F(z) = det(z) (N = n) is quasi-convex but it is not convex. To see that, let g(x) be a function of class C?(R",R”), and let w be the differential form det(Dg)dxy Adar N+++ A dity - We have w = dgi Adgg \---Ndgn = d(gi Adg2 A+: \dgn) (5.9) since dd = 0, and therefore ff ser Da)dey Nea Ao Aden -{ Gd nde In lon In particular, the integral on the left-hand side is zero whenever one among the functions 91,...,gn is zero on 89. Moreover, if two functions 9 and h coincide in a neighborhood of 82, we have ii det(Dg) dz = [ det(Dh) de. ln n The same result (without assuming N = n) holds if we substitute if the determinant any minor of the matrix Dg. In particular, if y is a function of class C2(@), the functions u(x) = (zo,2) and u+y have the some value in a neighborhood of 89, and hence f det(Du + Dg) dx = f det(2o + Dy) dx = |0| det(z0) . In In By approximation with C? functions we can easily see that the above relation holds for an arbitrary function y € W,'"(2), so that the function det(z) is quasi-convex.? More generally, let z be a n x m matrix, and let M(z) be the vector whose components are the minors of the matrix 2. If g(M) is a convex Actually, the value of the functional F(u) = [det(Du) dx depends only on the boundary values of u, and therefore it is the same for functions having the same values on 9. Consequently, the EULER equation for F is identically satisfied. Functionals of this kind are called null Lagrangian. Quasi-Convex Functionals a7 function of M, the function F(z) = 9(M()) is called polyconver. We have Proposition 5.1 A polyconver function is quasi-conver. Proof. If g(w) is a convex function and w(z) a summable function in 9, we have the JENSEN inequality ff atotey ae > 9 (f tear) (6.10) Setting u = (zo,2), we have faonco +Dy)) dz > 9 (fe + Dea) : ln = a(f Meo) 4c) = 9040) from which the result follows at once. 7 A last and weakest form of convexity is the so-called rank-one convezity.> Definition 5.3 A function F(x,u,z) defined in Qx RN x R™ is rank- one convex if for every (0, uo, 20) € 2 x RN x R™ the function 9(E,n) = F(x0, uo, 20 + € ® 7) is separately conver in €€ R” andne RN. Proposition 5.2 A continuous quasi-convex function is rank-one convex. Proof. We can suppose that F depends only on z. Assume first that F is of class C?, Since F is quasi-convex, the affine functions minimize the functional F, and therefore if u(x) = 20 is an affine function and if 3In his book [3], modifying a terminology that he himself had previously introduced in (2], Monrey calls quasi-convexity the rank-one convexity, and strong quasi-convexity what, according to current usage, we have called quasi-convexity. 148, Direct Methods in the Calculus of Variations ¢ € C4(0,R%), the function G(t) = Flu t ty) has a minimum for t = 0. We have therefore G’(0) = 0 and G"(0) > 0, and hence erat ) Dey" Dy! de > 0 (5.11) In zg02h oe . : for every y € O2(9, RY). Setting y = A+ ip, writing (5.11) for A and yz and summing, we get @F Re | ——(20) Dip*Dj" dz > 0. 5.12) I gagarpoo Pv” DP? de > (6.2) ‘We choose now y = ne'7\*)4(x), where € € R", 1 € RN, and yisa fanction in Cg°(,R). From (5.12) we obtain OF 2 2 1 + DyyDjy\ dx > 0. Loxeae™ a lr ues? + DeryDjy\ de > Dividing by 7? and letting 7 go to infinity we get OF fsa colegiado 2 0 In azgo28 for every 7 € C°(), and in conclusion OF aa (2o)eeinon? > 0. (6.13) axgazs The last inequality, that carries the name of the LEGENDRE-HADAMARD condition, implies the separate convexity in € and 7. Suppose now that only F is continuous, and let Fe) = f Fle whew) aw be the mollified function of F. If # is a function in C4(), we have f[ Filzy + Dd) de = [ec f F(29 + D9 —w) de In In z J ecuyrte — w)|Q| dw = |Q|F.(z0) and hence F, is quasi-convex. Quasi-Conver Functionals 149 It follows from the above that F. is rank-one convex, and passing to the limit as € + 0 the same is true for F. a We have then a series of conditions of increasing generality, since con- vexity implies policonvexity (we can consider the single components of the gradient as rank-one minors), the latter in turn implies quasi-convexity, which finally implies rank-one convexity. Generally speaking, the opposite implications do not hold,* except in special situations. A first case, in which all the definitions above are equivalent, is when u is a scalar function, that is when the codomain has dimension N = 1. In fact in this case the LEGENDRE-HADAMARD condition reduces to OF 202; (20)€i; 20 and is equivalent to the convexity of the function F.5 ‘A second case of some importance, in which rank-one convexity implies quasi-convexity, is when the function F(z, u, z) is quadratic in z: F(z, u, 2) = Ads(x,u)2ez? . When F has the above form, setting for simplicity A = A(zo, uo), we have J F210, 20+ Dp)de = | (Azo,20) + (ADe,Dp)] dr In 7 since the remaining integrals are zero because y has compact support. To prove the quasi-convexity it will therefore suffice to show that [(aDe,Deyde >0 : for every p € C§°(2). More generally, assume that instead of (5.13) we have Ad En? > vie? In? (6.14) with v > 0. “For a more detailed discussion we refer to the final section of this chapter. SIf F is not regular, we can proceed as above, approximating with regular functions. We note that the same conclusion also holds when the domain has dimension n = 1. 150 Direct Methods in the Calculus of Variations In this case we can prove the following: Lemma 5.1 Let A be @ constant matriz, satisfying (5.14). For every CE Wh? we have [4% Dj cP. de> v [DCP ae Proof. Denoting by f(€) the FOURIER transform of a function f: 1) = Gage | Here de, we have Dy f(€) = itn f(€), and [peepee ae [ Deedee = | 6eiG@P a. On the other hand, if in (5.14) we allow the vector 1 to take complex values: * = 9% + in, we get AY, 6:6)? = AZsE6;[0% 0? + 127? + (0%? — ox) and hence Re AiZ,E:€5n°n? > viE)? inl. As a consequence, recalling PARSEVAL’s identity: / A9,DjCPD C2 dr = Re [4 AB 66;C2C de av feriirag=v fiocPas and the proof is concluded. a 5.2 First Semicontinuity Results ‘The results of the preceding section force us to ask whether the quasi- convexity is also sufficient for the lower semicontinuity in a topology weak enough to guarantee the existence of minima. Results in this direction are rather recent, and have been proved under increasingly less restrictive conditions. In this section we shall consider the relatively simpler case in which the integrand F depends only on z; the general case will be treated later. Quasi-Conver Functionals 151 ‘To begin, we need the following lemma: Lemma 5.2 Let F(z) be a rank-one conver function, such that IF) SeA+|zI)?, 420. (5.15) with p > 1. Then, |F(z) — F(w)| < c(A+ |z| + fwl)?7 22 - w]. (5.16) Moreover, if F has first derivatives, we have [Fil 1 we obtain F(2*)) — F(2) = Ga) < Git) Fe) + (2) — 2)) — Fe) = A+ [al + [wl cot Using (5.15) and remarking that |2(*)| < |z|+|w| and |2) + 4(2+) — 2)| < (A+ |2| + |wl), we get (2D) — F(2) < o(d+ fel + |w))P—? wl, from which, summing over k and then exchanging z with w, we deduce immediately (5.16). Now let v be a vector in R", with |v] = 1. From (5.16) we get F(z-+tv) — F(z t Se(A+|z + tv] +|zl)?-?, and (5.17) follows immediately letting t go to zero. ao 152 Direct Methods in the Calculus of Variations We can prove now the first semicontinuity theorem. Theorem 5.4 (MARCELLINI [1]) Let F(z) be a quasi-convex function, such that 0 < F(z) 1, Then, the functional Fluo) = [ F(Du) dr In is lower semicontinuous in the weak topology of W,i:?(2,R™). Proof. Let u, > u in W1?(Q), and assume first that u is an affine function, so that Du = z = constant. If we had u, = u on O2 (that is u, = u+ yp with gy € W2'?(Q)), then from the assumption of quasi- convexity we would get [Fuyae > Foain\= [ FDwas n A and the semicontinuity would follow immediately. The problem therefore consists of modifying suitably the functions u, near 8, in such a way that they take the value u on 92 without changing excessively the value of the functional. For that, let Mo CC Q and let L be an integer. We set R= }dist (%, 9M), and for i= 1,2,...,L we define a, {re Q: dist (2,9) < zn} ' Now choose functions J; € C}() such that OSMS1; Wel inM%1; [Dy < MD and set Vin = U+ Yi(ta —u)- ‘The functions vin —u belong to Wy’?(). From what we have said above it follows that [Poemar> Poaial= [ Fwd, a In Quasi-Conver Functionals 153 and hence f F(Du) dx < [ F(Du) de + F(Dun)de+ f F(Dun) dx A nn, M1 a < [ F(Du) de + i (Duyn) de + if F(Dup) de. 2-2 125A 2 (5.18) On the other hand Duin = (1 — Yi)Du + biDun + (un - u)D¥i, and therefore (L 7 [Dunl? 0 there exists a family Qi such that f Du-Zirae => f a. fy au ra Let now u, be a sequence converging weakly to u in Wh?(Q). For x € Q,, let vf (2) = ua(z) — u(z) + (2,2). When h tends to infinity, the sequence vu tends to (zi,2) in the weak topology of W1(Q,), and therefore, according to what we have just proved, we have liminf f F(Dol) de > ii F(x) dx hove JQ, 1d and the same inequality holds if we sum over i. On the other hand, using the preceding lemma and the HOLDER inequality, and summing over i, we obtain ff P(Dup) de — 2 iE ' F(Do}”) da] < eye ( + [Dup|P-? + [De P-*) |Du - zi|dz res a : < {rf (14 |Duyl? + ioral {Ef ipa sivas} Se. Ina similar way we prove that [Foo de Be F(a) dz u in W,,?(Q), we have from any 5 CCQ: F(u, ©) < liminf F(un,B) < Tim inf F(un, 2) and the theorem follows letting D tend to 2. o Remark 5.2 The conclusion of the theorem remains valid if we only assume that —c(1 + |z|") < F(z) < e(1 + |z|P) with r < p. In this case we must add to the right-hand side of (5.18) the quantity -[ F(Ds) de [ F(Dun) de i No 2-11 < ef (1+ |Dul + [Dual)" de. Inn The last integral can be estimated by : in 91°F ( [1+ |Du|+ Dun)ras)” 2. Everything else remains unchanged. Qo ‘The preceding result continues to hold when the function F depends also on ¢ and w. In this case the proof is considerably more complex, and we shall need some preliminary results, many of which are interesting in themselves, independently of their application to the semicontinuity. 5.3 The Quasi-Convex Envelope Definition 5.4 Let G(z) > 0 be a function defined in R™, and let 2 be an open set in R". We set _ {foe + Dela) de; p € oxen} a In It is easily seen that 99 is invariant under homothety. Actually, if 21 is an open set, homothetic to 2: Q; = co +A, and if y, € C#°(M1), setting (2) = A~Nps (arg + Az), we have y € CF°(M), and foes Dow d= f Ge + Dee) de, ln a from which immediately follows yo = Y,- 162 Direct Methods in the Calculus of Variations In particular, if ug is a minimizing sequence, that is if ¢, = F(uz)—inf F tends to zero, the corresponding sequence v is itself a minimizing sequence. Moreover, we have Flue) < Flw) + fd(oz, w) (5.32) for every w eV. o 5.5 Semicontinuity We can now prove the main result of this chapter. For that, we shall consider a CARATHEODORY function F(z, u, 2), defined in Qx RN x R™, and satisfying the conditions: ~a(z|" + lu") - A(e) < F(@,u,z) < 9(@,u)Q + [z!") (5.33) t > 1ifp>n). As for the with p>1,1 0 isa functions A and g, we shall assume that h CARATHEODORY function in @ x RN. Under these assumptions we shall prove the following semicontinuity result: Theorem 5.7 (AceRBI and Fusco [AF1]) Let F(z,u,z) be a quasi- conves function, satisfying the conditions (5.33). Then, the functional Flu) = f Flesu(e),Dula) de ln is lower semicontinuous in the weak topology of W,y?(2,R™). The proof of the theorem will be made in a series of steps. In the first place, we shall prove the theorem when the lower bound in (5.33) is replaced by F(a,u,z)>vlzP, v>0. (5.34) For i € N we call ¥,(t) a continuous function taking the values one for t i, and we set (|u|) if g(a,u) Si, #9;(\ul) g(@,u) HZ, U, a if g(x,u) >i. Moreover, we set F(a, 4,2) = m(z,u)F(@, u, 2) + (L— m(2,u)) 168 Direct Methods in the Calculus of Variations Let now uj — u in W,4?(Q), and let % CC @ be an open set with regular boundary. Since t < p*, the sequence u; will converge to u in the strong topology of L'(Q). We have therefore: limint | F(e,u;, Dus) dx n Vv timint [ P(e, Ds) de ~ fin [ {alu;|t + A(z) +.M} de — 2ce ath ie iv ff Flu Duyds ff Kolul + h(a) +My de ~ Dee No ‘No 2f F(z, u, Du) de — 2ce ‘No and the conclusion follows letting first ¢ + 0, and then letting % tend to. Remark 5.6 The preceding theorem does not hold if either r = p, or t = p*. In the first case, MURAT and TARTAR [1] have shown that the functional i det(Du) dx fy (n= N = 2) is not continuous in the weak topology of W1? (note that the functions det(z) and — det(z) are both quasi-convex). In the second, it will be sufficient to consider the functional [our lw" )ae In and a sequence uy converging strongly to u in W1?, but not in LP". 0 5.6 Coerciveness and Existence The above results can be applied to prove the existence of minima of functionals. As we said in the preceding chapter, once the semicontinuity has been proved, the existence of minima under suitable conditions will depend essentially on the coerciveness of the functional under discussion. In the preceding chapter we have seen that for the DIRICHLET problem the coerciveness follows from estimates of the type lal? - B(e)|ul? - g(x) < F(@,u,z) < Lle|P + b()|ul7 +9(2). (6.86)

You might also like